[Whisper] 🚨 Fix whisper decoding 🚨 #34135
eustlb
merged
75 commits into
huggingface:main
from
eustlb:fix-whisper-decoding
do not remove decoder_input_ids for the first segment
4c7be8c1
do not remove eos token in generate_with_fallback
805c6887
when removing padding tokens, do not remove eos token
271014fd
remove eos token in generate (and not in generate_with_fallback!)
d9a91b3e
reconciliate short-from/ long-form behavior
b9d0ec14
correct avg_logprobs calculation
f9d5fdd8
handle eos token in segments
f111aa34
handle decoder_input_ids and eos token in _prepare_decoder_input_ids
34d690aa
fix incorrect time precision
16f768b3
always remove eos token
cf18e0a3
always remove decoder_input_ids
ec7cd58b
no need to handle decoder_inputs_ids and eos token
15309306
no need to remove decoder_input_ids
67865dd1
no need to handle eos token
eb107d9c
fix num_beams in _retrieve_logit_processors
78819284
remove todo unconsistency
031ace6f
no need to add eos token
eaaeec60
last_timestamp_pos should indeed be timestamp token pos
cdd51447
patch generate to enable compatibility with GenerationTesterMixin tests
544f21b6
adapt test_generate_continue_from_past_key_values
70ffdb30
adapt test_prompt_lookup_decoding_matches_greedy_search
79347e9f
eustlb
changed the title
Merge branch 'main' into fix-whisper-decoding
c01ef124
Merge branch 'main' into fix-whisper-decoding
d03f87cf
adapt generic GenerationMixin tests to whisper's generate
1387919f
fix speculative decoding
ad262c66
fix
a4c2b448
Merge branch 'main' into fix-whisper-decoding
17877f3f
Merge branch 'main' into fix-whisper-decoding
7b2046a8
Merge branch 'main' into fix-whisper-decoding
c9a209fd
[run-slow] whisper
34d5257e
change HF_HUB_TOKEN for require_read_token
c760d095
[run-slow] whisper
e7c4b85a
Merge branch 'main' into fix-whisper-decoding
40bcdccf
prioritize kwargs over generation_config
783c2f1b
remove unnecessary args
88d2d38b
[run-slow] whisper
3a8f7390
update tests
93cf1b73
Merge branch 'main' into fix-whisper-decoding
2afb05cf
[run-slow] whisper
ecbe56b3
add comment
c9b36c3a
update test
14c280a1
Merge branch 'main' into fix-whisper-decoding
78ed67eb
[run-slow] whisper
34adc990
update test + revert require_read_token
58f0526e
docstring updates
367006d7
revert tokenizer decode args change
2f19263b
Merge branch 'main' into fix-whisper-decoding
9c6fe558
do not use a patch + docstring updates
3ebd12f5
Merge branch 'main' into fix-whisper-decoding
b9ba08a2
[run-slow] whisper
24d93990
make
9fe30114
[run-slow] whisper
3aae6b80
ylacombe
commented
on 2024-12-10
eustlb
changed the title
add a flag to force unique call to generate
0ea240fa
test update
0960a525
[run-slow] whisper
0870ac7a
add force_unique_generate_call arg
00d37e82
do not use a patch
cf75ea37
correct the timestamps for the pad tokens
2cb638d3
docstring update
66290cb2
docstring update
b41e2c13
docstring update
221bae15
Merge branch 'main' into fix-whisper-decoding
98dc7f1f
upodate TF tests
0b6687cf
add require_read_token
ab910f74
[run-slow] whisper
692bf147
test reset dynamo
a15aa4a4
Merge branch 'main' into fix-whisper-decoding
0011cec5
[run-slow] whisper
41df6ca0
fix
aaecea40
[run-slow] whisper
58d0b90c
ArthurZucker
approved these changes
on 2024-12-17
avoid iterating twice on current_segments
a1f4e431
[run-slow] whisper
671b0795
Merge branch 'main' into fix-whisper-decoding
3d21ed8e
[run-slow] whisper
dc6fbd1e
Merge branch 'main' into fix-whisper-decoding
8b7a2e83
Login to write a write a comment.
Login via GitHub